Login
Computer Science @ Rochester
Monday, August 30, 2004
11:00 AM
CSB 209
Kevin Knight
USC Information Sciences Inst.
Machine Translation = Automata Theory + Probability Theory + Linguistics
Recently, machine translation (MT) systems have become much more accurate. A major reason is that machines now gather translation knowledge autonomously, combing through large amounts of human-translated material available on the web. What the systems learn are essentially finite-state Markov models---target sub-strings are substituted for source sub-strings, followed by some re-ordering. A serious weakness, however, is that this kind of model can only support very weak linguistic transformations, and the trained models do not yet lead to reliably high-quality MT.

Over the past three years, many new probabilistic tree-based (versus string-based) models have been designed and tested on many natural language applications, including MT. Most of these models turn out to be instances of tree transducers, a formal automata model first described by W. Rounds and J. Thatcher in 1970. These automata open up new opportunities for us to marry deeper representations, automata theory, and machine learning. This talk will cover new learning algorithms for tree automata, together with experiments in machine translation.

This is joint work with Jonathan Graehl (USC/ISI), Daniel Marcu (USC/ISI), and Dan Gildea (Rochester).